Sep 29

QP 26 - The Right Patient, the Right Treatment: Machine Learning for Stratification and Prediction

1285 - Deep Learning-Enabled IMPT Treatment Planning Using a Current Clinical Setup: Principles and Validation Toward Online Adaptive Proton Therapy

05:45pm - 05:50pm ET

Room 204

Presenter(s)

Minglei Kang, PhD - University of Wisconsin, Madison, WI

B. Pang¹, H. Li², M. Kang³, L. Zhang⁴, Y. Ma⁵, M. Chen⁶, K. Yang¹, and Z. Yang¹; ¹Cancer Center, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China, ²Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University School of Medicine, Baltimore, MD, ³Department of Human Oncology, University of Wisconsin-Madison, Madison, WI, ⁴Medical Artificial Intelligence Lab, The First Hospital of Hebei Medical University, Hebei Medical University, Shijiazhuang, China, ⁵Department of Radiation Oncology, The First Affiliated Hospital of Zhengzhou University, Henan, China, ⁶Department of Radiation Oncology, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China

Purpose/Objective(s): The optimization of proton therapy plans is a highly time- and labor-intensive process. Moreover, to prevent the effectiveness of the initial plan from being compromised during treatment, adaptive proton therapy strategies have become increasingly important, leading to a growing demand for fast and automated adaptive proton therapy treatment planning workflows.In this study, we investigate the feasibility of predicting intensity modulated proton treatment (IMPT) plan weights using a deep learning model to achieve a faster and more general automated adaptive proton therapy treatment planning workflow.

Materials/Methods: The 3D Attention U-Net network was employed to generate three-dimensional weight map, with the masked mean absolute error (MAE) used as the loss function. The model input was a four-channel 4D tensor comprising the robust optimization plan dose, reference dose, CT, and spot mask, with the planned spot weights serving as the labels. To evaluate the model generalizability, training and testing were performed on prostate, brain, and lung cases. The total workflow was validated in the Eclipse^® treatment planning system and the plan was delivered by the ProBeam 360° proton therapy system.

Results: Regarding the errors between the predicted and planned spot weights for all test samples, the prostate model achieved a mean relative error of less than 10%within the main distribution range of the planned weights, followed by the brain model with errors generally below 20%, and finally the lung model with errors generally below 40%. For the predicted dose distributions calculated based on the predicted weights, under the (2%, 2 mm) criterion and 10% minimum dose threshold, the prostate and brain models achieved gamma passing rates of 99.73±0.32% and 98.95±0.86%, respectively. The lung model, however, exhibited a gamma passing rate of 92.75±3.52% due to the difficulty in predicting weight variations in regions with large differences in tissue density. Under gamma analysis criteria (3%, 2mm), the plans delivered by the ProBeam 360° proton therapy system across different anatomical sites satisfied the QA acceptance criteria.

Conclusion: This study is the first to introduce a deep learning–based method to directly predict proton robust plan weights and prove the clinical feasibility by importing and delivering it with the clinical ProBeam360° proton therapy system. This enables a fast and automated treatment planning workflow and further advancing the development of automated online adaptive proton therapy.